1 / 10

Chapter 7

Chapter 7. System Model – typical assumptions underlying the study of distributed deadlock detection Only reusable resources, only exclusive access, single copy of resource in system. Besides deadlocks due to resources, we can have communication deadlocks Strategies

tracy
Download Presentation

Chapter 7

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Chapter 7 • System Model – typical assumptions underlying the study of distributed deadlock detection • Only reusable resources, only exclusive access, single copy of resource in system. • Besides deadlocks due to resources, we can have communication deadlocks • Strategies • Prevention – can cause further problems. Consider one shot allocation where resources and requesters are on different sites. • Avoidance – global state needs to be maintained, safe state checks need to be mutually exclusive.

  2. Detection • Issues • Detection • Deadlocks should be detected in finite time. • No false positives (phantom deadlocks) • Realize that states are not coherent. • Resolution • Clean up the information upon rollback. • Control Organizations • Centralized • Distributed • Hierarchical

  3. Distributed Deadlock Detection • Basic Centralized. • All requests and release messages are sent to a designated site, which maintains a global WF Graph • Problems – bottleneck, single POF, phantom deadlocks • Ho-Ramamoorthy 2 Phase • Each site maintains a status table – resources locked and resources being waited upon. The central site periodically requests this table, constructs a global WFG, and searches for cycles. If a cycle is found, it requests the tables again, and constructs WFG from those transactions that are common to both tables. If a cycle is still detected, then a deadlock is declared.

  4. Ho-Ramamoorthy 1 phase • Each site maintains 2 tables, one for resources (transactions that have locked a resource) and one for process status (resources locked/waited). These tables requested periodially by central site, and WFG constructed using those entries in the resource table which have corresponding entry in process table. • Distributed Algorithms • Path pushing: WFG constructed by disseminating dependency sequences • Edge chasing: process sends out probes. A blocked process receiving probes circulates it along its outgoing dependency edges • Diffusion: queries are diffused (sucessively propagated) and reflected • Global State Detection

  5. Obermack’s Algo. • Path pushing approach deals with transactions. Each transaction may have sub transactions, but they execute sequentially. Transactions are totally ordered. • Each site waits for deadlock related information (paths) from other sites. It abstracts the nonlocal portion of the WFG with a single node called EX. • It combines this with its own WFG. It then detects cycles and breaks those which do not contain EX. • For all cycles involving EX, the string indicating the cycle (EX-T1-T2-EX) is sent to all sites which have subtransactions of T2 waiting to recv a message from the subtransaction of T2 at this site. • Problem – this algorithm can detect phantom deadlocks. Needs n(n-1)/2 messages of O(n) size and detects in linear time.

  6. Chandy-Misra- Haas • Edge chasing algorithm based on the AND model. • A process Pj is dependent on Pk if there is a sequence Pj, Pi1….Pin,Pk such that all process but Pk are blocked, and each process except Pj has something that is needed by its predecessor. • Locally dependent • If Pi is locally dependent on itself, then we have a deadlock. Otherwise • Forall Pj, Pk such that Pi locally depends on Pj and Pj is waiting(not locally) on Pk, send probe(i,j,k) to Pk.

  7. On receiving probe(i,j,k) • If ( Pk is deadlocked && ! dependentk(i) && Pk has not replied to all requests of Pj ) • Dependentk(i) = true. • If (k == i) • Then Pi is deadlocked • Else Forall Pm, Pn such that Pk locally depends on Pm and Pm is dependent (not locally) on Pn, send probe(i,m,n) to Pk. • Sends 1 proble message on each edge of WFG, so m(n-1)/2 messages for a deadlock with m processes over n sites. Size is fixed, and detection time is linear in number of sites

  8. Diffusion Based Algorithm • Works for OR request model • Initiation: • A blocked process i sends query(i,i,j) to all Pj in its dependent set; numi(i) = |DSi| , waiti(i) = true; • When a blocked process Pk recvs query (i,j,k) • If this is engaging query, send query(i,k,m) to all processes in its dependent set, and set numk(i) and waitk(i) • Else if waitk(i) then send reply(i,k,j) • When Pk gets reply(i,j,k) • If waitk(i) • Decrement numk(i), if it becomes 0 then • If k == I then deadlock else reply(i,k,m) to the process which sent the engaging query.

  9. Heirarchical Algorithms • Menasce-Muntz • Resources are managed by nodes that form the “leaves” of a tree. They maintain TWF/WFGs corresponding to the resources they manage. • Several leaf controllers have a single parent, and so on in a tree fashion. Each non-leaf controller maintains WFG which is union of child WFGs. Changes are propagated upwards, and deadlocks detected on the way • Hierarchical Ho-Ramamoorthy • Sites split into disjoint clusters. • Each cluster has its own control site. There is also a central control site.

  10. Issues • Formal methods to prove correctness • Performance metrics • No of messages ? Message size? Time to detect ? Storage overhead ? Computation overhead ? • Resolution – basically aborting a process • How does a process know which others are involved in a deadlock ? • Can two process detect the same deadlock simultaneously ? • Use Priorities! • Rollback – release resources, clean up graph • Phantom Deadlocks.

More Related